Exploring Data Science, Project Development, Creative Writing, and More
Latest Research
Mechanistic interpretability papers, traced from attention heads down to SAE features.
The Frequency Prior Series
How GPT-2 encodes and yields to training-frequency priors — traced from attention heads down to SAE features. Every result from real-model inference, now explorable in an interactive companion app.
Featured Apps
Three standalone interactive apps — open them in any browser, no install required.
Hypersphere Explorer
Live Three.js visualization of spheres, cylinders, and 4D hyperspheres with KaTeX math that updates in real time as you drag sliders. Three guided tours walk through Archimedes' 2:3 ratio, 4D slicing, and the curse of dimensionality.
Launch →Chroma Echo
Audio-reactive particle canvas where mic input drives burst patterns and a built-in keyboard (A–K) plays synth tones that spawn particles mapped to pitch. Five choreography patterns — firework, vortex, spiral, rain, shockwave — fire from a single keystroke.
Launch →Mech Interp Viz
Research-grade mechanistic interpretability workspace for transformer models. Visualize attention heads and MLP blocks across every layer, annotate components with importance levels and tags, and export findings as structured data.
Launch →Get In Touch
Have a question, want to collaborate, or just want to say hello? I'd love to hear from you.